Jaegi Hwang  Seongju Kang  Kwangsue Chung

10.5626/JOK.2022.49.8.627

Deep Learning Object Detection GCN Attention

Feature Pyramid Network (FPN) is a feature map fusion technique used to solve the multi-scale problem of object detection. However, since FPN performs feature map fusion by focusing on adjacent resolutions, there is a problem in that semantic information included in non-adjacent layers is diluted. This paper, proposes a graph convolution network (GCN)-based feature map fusion technique for multi-scale object detection. The proposed GCN-based method dynamically fuses feature map information of all layers according to learnable adjacency matrix weights. The adjacency matrix weight is generated based on the multi-scale attention mechanism to adaptively reflect the scale information of the object. The feature map fusion process is performed through a matrix multiplication operation between adjacency matrix and a feature node matrix. The performance of the proposed method was verified by showing that it improves the multi-scale object detection performance in the PASCAL-VOC benchmark dataset compared to the existing FPN method.